Literary Translation: Why Machine Translation Can't Capture Style

The Misconception: Machines Can Translate Literature

It's easy to assume that with neural machine translation (NMT) systems like Google Translate and DeepL, literary translation is just a matter of running a text through a tool and polishing the output. This is wrong. While NMT has made impressive strides, it fundamentally misconstrues what literary translation is about. Literary translation is not about transferring meaning from one language to another; it's about recreating an aesthetic experience, capturing voice, rhythm, and cultural nuance in a way that resonates with a new audience. Machines, as we'll see, are not equipped for that task.

Consider this: when Google announced its multilingual neural machine translation system in November 2016, it was already translating over 140 billion words every day across 103 languages (Zero-Shot Translation, Google Research). That's an enormous volume, but almost all of it is informational: news, manuals, emails, social media. Very little of it is literature. And there's a reason for that. The very nature of literary prose—its ambiguity, its subtext, its deliberate use of style—resists the kind of literalistic processing that even the best NMT performs.

So, what is the real state of literary translation in the age of neural networks? To answer that, we need to ask a sharper question: Should translators use machine translation as a starting point for literary work, or does it do more harm than good?

What Machine Translation Actually Does Well

Machine translation has come a long way since the early days. The field began with Warren Weaver's 1949 memorandum proposing the use of digital computers for translation (Timeline of Machine Translation, Wikipedia). After a rocky period—the 1966 ALPAC report was so skeptical that it led to US government funding cuts (ALPAC, Wikipedia)—research persisted, eventually giving rise to statistical machine translation (SMT) in the 1990s. SMT, with its IBM models, worked by analyzing large parallel corpora, but it was phrase-based and often stilted. Then came the Transformer architecture in 2017, which, as the paper "Attention Is All You Need" showed, could achieve a BLEU score of 28.4 on English-to-German translation, surpassing previous best results (Attention Is All You Need, NeurIPS 2017). This paved the way for modern NMT.

Today's NMT systems are remarkably good at producing fluent, accurate translations for informational texts. For instance, Microsoft's Translator, when it switched to neural networks in 2016, immediately improved speech translation quality (Microsoft Translator, Microsoft Learn). Amazon followed suit with its own neural service in 2017 (Amazon Translate, AWS News Blog). And on the evaluation front, metrics like COMET and BLEURT now correlate better with human judgment than the older BLEU metric (IBM Research). So, if you're translating a technical manual or a legal document, machine translation can save a lot of time.

But here's the rub: literary translation is not informational. It's expressive. And that's where machines hit a wall.

Why Style and Voice Defeat Neural Networks

Literary translation is a craft that involves interpreting not just the denotative meaning of words but their connotative weight, their sound, their allusions. A machine, trained on large corpora, can guess at the most probable translation of a phrase, but it cannot make the kind of aesthetic judgment that a human translator makes when choosing between two equally "accurate" renderings based on the rhythm of a sentence or the tone of a character's speech.

Take poetry, for example. A poem's meaning is often inseparable from its form—its rhyme scheme, meter, line breaks. A machine translation that renders a poem into prose is not a translation; it's a paraphrase at best, and a betrayal at worst. Even in prose, authors like James Joyce or Gabriel García Márquez play with language in ways that resist literal transfer. The translator must recreate those effects in the target language, often making extensive changes that a machine would never consider.

One might argue that a human post-editor could take a rough machine translation and refine it into something literary. But post-editing is a specific skill, and ISO 18587:2017 defines standards for post-editing machine translation output (ISO 18587, ISO). Yet even full post-editing, which requires the editor to bring the translation up to "human" quality, is a tall order. The post-editor would essentially have to rewrite the entire text, and at that point, why not just translate from scratch? A machine translation can even be a distraction, luring the translator into literalisms and awkward constructions that they'd never have produced otherwise.

Moreover, the evaluation metrics used in machine translation reveal its limitations. BLEU, which compares n-gram overlaps with human references, is notoriously poor at capturing stylistic quality (IBM Research). Newer metrics like METEOR, which include synonym and paraphrase matching (METEOR, CMU), and chrF, which is character-based and language-independent (chrF, ACL Anthology), are better but still fundamentally measure surface similarity. None of them can tell you if a translation is beautiful.

The Role of Human Translators in Literary Translation

So, where does that leave us? Is there no place for machine translation in literary work? Actually, there is a narrow but useful role: as a tool in the very early stages of exploring a text, particularly for translators working on a piece in a language they're less fluent in. For instance, a translator might use machine translation to get a rough sense of a novel's plot before reading it in the original. But that's a far cry from using it as a drafting tool.

The language services industry is already feeling the pressure. Market estimates for 2023 show the industry at $49.68 billion according to CSA Research, while Slator puts it at $27.03 billion, depending on scope (CSA Research; Slator). That's a lot of money, but it's worth noting that the industry includes everything from subtitling to software localization. Literary translation is a tiny sliver, and it's unlikely to be automated away.

Institutional translation services, like those of the European Commission, which produced about 2.6 million pages in 2022, rely heavily on machine translation and post-editing for their vast documentation (European Commission). But they also employ human translators to ensure quality. For literary translation, the standard is even higher: you need someone who can capture the author's voice, preserve cultural allusions, and make the text read as if it were written in the target language.

Consider the example of a translator working on a Japanese novel. The author might use a particular dialect or level of politeness that conveys social hierarchy. A machine translation will likely flatten that into standard target language, losing the nuance. A human translator, on the other hand, might choose to use a regional dialect or colloquialisms in the target language to convey a similar effect. That's a creative decision that no algorithm can make.

Furthermore, the field of literary translation is not just about language; it's about culture. A translator must decide how to handle references to local customs, foods, or historical events. Sometimes they'll keep the original term and add a footnote; other times they'll find an equivalent in the target culture. These decisions require deep cultural knowledge and an understanding of the reader's expectations.

So, what is the single best move for a literary translator today? Embrace machine translation as a tool for reference, not as a collaborator. Use it to quickly look up a phrase or to check your understanding of a passage, but never let it draft your prose. Instead, invest in your own craft: read widely in both your source and target languages, study the techniques of other translators, and develop your own sense of style.

The truth is, literary translation is an art, and machines are not artists. They are, at best, efficient clerks. As the great translator Edith Grossman once said, "Translation is not a mechanical process; it is a creative act." That creativity is what makes literary translation worth doing, and it's what will keep human translators in demand no matter how sophisticated machine translation becomes.

Bottom Line

If you're translating literature, do not use machine translation as a starting point. It will only lead you astray. Instead, translate from the original text, using your own linguistic and cultural expertise to recreate the author's voice in your target language. Machine translation is a useful reference tool, but it cannot replace the human touch that literary translation demands.

Sources

• "Machine Translation" - Wikipedia - https://en.wikipedia.org/wiki/Machine_translation
• "Attention Is All You Need" - NeurIPS 2017 - https://arxiv.org/abs/1706.03762
• "Zero-Shot Translation" - Google Research - https://research.google/blog/zero-shot-translation-with-googles-multilingual-neural-machine-translation-system/
• "ISO 18587:2017" - ISO - https://www.iso.org/standard/62970.html
• "METEOR" - CMU - http://www.cs.cmu.edu/~alavie/METEOR/
• "chrF" - ACL Anthology - https://aclanthology.org/W15-3049/